ESCO1 and CTCF enable formation of long chromatin loops by protecting cohesinSTAG1 from WAPL

Eukaryotic genomes are folded into loops. It is thought that these are formed by cohesin complexes via extrusion, either until loop expansion is arrested by CTCF or until cohesin is removed from DNA by WAPL. Although WAPL limits cohesin’s chromatin residence time to minutes, it has been reported that some loops exist for hours. How these loops can persist is unknown. We show that during G1-phase, mammalian cells contain acetylated cohesinSTAG1 which binds chromatin for hours, whereas cohesinSTAG2 binds chromatin for minutes. Our results indicate that CTCF and the acetyltransferase ESCO1 protect a subset of cohesinSTAG1 complexes from WAPL, thereby enable formation of long and presumably long-lived loops, and that ESCO1, like CTCF, contributes to boundary formation in chromatin looping. Our data are consistent with a model of nested loop extrusion, in which acetylated cohesinSTAG1 forms stable loops between CTCF sites, demarcating the boundaries of more transient cohesinSTAG2 extrusion activity.


Introduction
In eukaryotic interphase cells, cohesin complexes are essential for the formation and maintenance of numerous long-range chromatin cis-interactions Hadjur et al., 2009;Nativio et al., 2009;Rao et al., 2017;Schwarzer et al., 2017;Wutz et al., 2017). These are thought to have both structural and regulatory functions, in the latter case by contributing to recombination and gene regulation (reviewed in Lin et al., 2018;Merkenschlager and Nora, 2016). Hi-C experiments have revealed that chromatin interactions are either enriched in genomic regions called topologically associating domains (TADs) or appear as more pronounced localized interactions which in Hi-C maps are visible as 'dots' or on the edge of TADs as 'corner peaks'. In Hi-C maps only these dots and corner peaks are referred to as loops, even though TADs are also thought to be formed by looping of chromatin (Dixon et al., 2012;Nora et al., 2012;Rao et al., 2014).
Most chromatin interactions that are mediated by cohesin are anchored at genomic sites that are bound by the insulator protein CTCF (Dixon et al., 2012;Nora et al., 2012;Rao et al., 2014) with which cohesin co-localizes genome-wide (Parelho et al., 2008;Wendt et al., 2008). Even though the sites at which long-range interactions are anchored can be hundreds of kilobases or even Megabase pairs (Mb) apart in the linear genome, the CTCF consensus motifs that are found at these sites are typically oriented towards each other, a phenomenon known as 'the CTCF convergence rule' (de Wit et al., 2015;Rao et al., 2014;Vietri Rudan et al., 2015). How cohesin and CTCF generate chromatin interactions in cells is unknown, but an attractive hypothesis posits that cohesin acts by extruding loops of genomic DNA until it encounters convergently oriented CTCF sites (Fudenberg et al., 2016;Sanborn et al., 2015). This hypothesis is supported by recent single-molecule experiments which have shown that in vitro cohesin forms DNA loops by extrusion in a manner that depends on NIPBL-MAU2 Kim et al., 2019). NIPBL-MAU2's main function was so far thought to be the loading of cohesin onto DNA (Ciosk et al., 2000) but the experiments by Davidson et al. indicate that NIPBL-MAU2 is also part of the active cohesin holo-enzyme that mediates loop extrusion . Loops of DNA can also be extruded by condensin (Ganji et al., 2018), which like cohesin belongs to the family of 'structural maintenance of chromosomes' (SMC) family of complexes (Strunnikov et al., 1993), and which forms DNA loops in mitosis.
Cohesin is a protein complex composed of multiple subunits. Three of these, SMC1, SMC3 and SCC1 (also known as RAD21 and Mcd1) form tri-partite rings, which in replicating cells are thought to entrap newly synthesised DNA molecules to mediate sister chromatid cohesion (Haering et al., 2008). During quiescence (G0) and G1, cohesin is dynamically released from chromatin via the activity of the protein WAPL and has a mean chromatin residence time of 8-25 min (Gerlich et al., 2006;Kueng et al., 2006;Tedeschi et al., 2013). Despite cohesin's dynamic interaction with chromatin, recent experiments suggest that some chromatin loops can persist over a significantly longer timescale of several hours (Vian et al., 2018). How this can occur is not understood. It is possible that chromatin interactions are maintained after cohesin has been unloaded, but the observation that experimentally induced degradation of cohesin's SCC1 subunit leads to disappearance of most TADs and loops within 15 min argues against this possibility (Wutz et al., 2017). More plausible scenarios are that long-lived chromatin interactions are maintained by multiple short-lived cohesin complexes, or that cohesin can be protected from WAPL so that it can persist on chromatin for longer periods of time.
Precedence for the regulatability of cohesin's residence times on chromatin comes from studies of proliferating somatic cells and from meiotic cells. During S phase of somatic mammalian cells, around half of all cohesin complexes become protected from WAPL via an incompletely understood mechanism that depends on acetylation of cohesin's SMC3 subunit (Rolef Ben-Shahar et al., 2008;Unal et al., 2008) by the acetyltransferases ESCO1 and ESCO2 and on the subsequent recruitment to cohesin of the protein sororin (Ladurner et al., 2016;Lafont et al., 2010;Nishiyama et al., 2010). This protection from WAPL increases cohesin's chromatin residence time to many hours and enables cohesive cohesin complexes to maintain sister chromatid cohesion from S phase until the subsequent mitosis (Gerlich et al., 2006;Schmitz et al., 2007). An even more dramatic prolongation of cohesin's residence time is thought to exist in mammalian oocytes. In these cells, a meiotic form of cohesin, cohesin REC8 , establishes cohesion during pre-meiotic S phase already before birth and then maintains it, depending on the species, for months or years until meiosis is completed during oocyte maturation cycles after puberty (Burkhardt et al., 2016;Tachibana-Konwalski et al., 2010).
Long residence times of cohesin on chromatin can also be experimentally induced by depletion of WAPL. This leads to re-localisation of cohesin from a diffuse nuclear pattern into axial chromosomal domains termed vermicelli, which are thought to represent the base of chromatin interactions (Tedeschi et al., 2013). This is accompanied by an increase in the number of long DNA loops and by chromatin compaction Haarhuis et al., 2017;Wutz et al., 2017). These findings imply that the residence time of cohesin on chromatin can be regulated in post-replicative cells in order to maintain cohesion over long periods of time, and they show that prolonging cohesin's residence time experimentally can have major effects on genome organization (but it is important to note that the stabilization of cohesive cohesin on chromatin during S and G2 phase does not detectably alter genome architecture, which implies that cohesion and chromatin looping are mediated by distinct populations of cohesin; Holzmann et al., 2019 and references therein). However, to date, a cohesin population with long chromatin residence times has not been identified in G1 phase, where some chromatin interactions have been reported to exist for hours (Vian et al., 2018).
In mammalian somatic cells, SCC1 associates with the subunits STAG1 or STAG2 to form two distinct tetrameric cohesin core complexes (Losada et al., 2000;Sumara et al., 2000). Although a recent study reported differences in the chromatin localization patterns for STAG1 and STAG2 and their contributions to chromatin organization (Kojic et al., 2018), it is unclear if these two forms of cohesin exhibit different residence times and incompletely understood whether they play distinct roles in loop formation. Here, we show that cohesin STAG1 and cohesin STAG2 do indeed display functional differences, both with respect to their residence times and in their ability to structure chromatin. We find that cohesin STAG1 complexes are more highly acetylated, interact more stably with CTCF, form larger chromatin loops, and have a longer residence time on chromatin than cohesin-STAG2 complexes, consistent with the proposed existence of long-lived chromatin loops. This stabilization of cohesin STAG1 depends on CTCF and ESCO1. Depletion of ESCO1 also decreases the insulation between TADs, in a manner like that observed following CTCF depletion. Furthermore, we find that both proteins are important for cohesin acetylation in G1. These results indicate that ESCO1 and CTCF function together to regulate cohesin's chromatin organization activity. Our results underline that precise regulation of cohesin's residence time is key to how cells organize their genomes. They may also be of relevance for understanding the aetiology of human cancers, in which STAG2 expression is often lost (Lawrence et al., 2014;Leiserson et al., 2015;Solomon et al., 2011).

ESCO1 preferentially acetylates cohesin STAG1 during G1 phase
Acetylation of cohesin's SMC3 subunit during S phase is known to stabilize cohesive cohesin complexes on chromatin (Ladurner et al., 2016), but acetylated SMC3 (SMC3ac) can also be detected in quiescent (G0) cells and in cells in G1 phase, where no cohesive cohesin exists (Alomer et al., 2017;Busslinger et al., 2017;Minamino et al., 2015;Rahman et al., 2015;Whelan et al., 2012). To address possible functions of SMC3 acetylation in G1 phase we first analyzed which cohesin complexes are modified during the cell cycle by ESCO1 and ESCO2. In this and subsequent experiments we detected acetylated SMC3 by using monoclonal antibodies that specifically recognize SMC3 which is acetylated either singly on K106 or doubly on K105 and K106 (Nishiyama et al., 2010). Immunoblot analyses of chromatin fractions isolated from synchronized HeLa cells confirmed that SMC3 acetylation is detectable throughout the cell cycle, as is ESCO1, whereas ESCO2 expression is confined to S phase (Figure 1-figure supplement 1A; lanes 2-5). As predicted from these results, depletion of ESCO1 by RNA interference (RNAi) reduced SMC3 acetylation in G1, whereas depletion of ESCO2 had little effect in this cell cycle phase (Figure 1-figure supplement 1B; note that contrary to the situation in G1, in G2 both ESCO1 and ESCO2 contribute to SMC3 acetylation; see also Nishiyama et al. (2010). As reported by Minamino et al. (2015), cohesin acetylation in G1 phase is therefore predominantly mediated by ESCO1.
Unexpectedly, however, we found that cohesin STAG1 and cohesin STAG2 were acetylated to different extents in G1. When cohesin STAG1 and cohesin STAG2 were isolated from G1 chromatin fractions by immunoprecipitation using antibodies that recognize the endogenous proteins, more acetylated SMC3 was detected in the former than the latter sample ( Figure 1A; compare lanes 4 and 6). These samples had been normalized to SMC3 to account for the three-fold higher abundance of cohesin-STAG2 relative to cohesin STAG1 on chromatin (Holzmann et al., 2019) but similar results were also obtained when STAG1 and STAG2 immunoprecipitates from the same number of cells were   Figure 1. ESCO1 preferentially acetylates cohesin STAG1 during G1 phase. (A) Immunoblot analysis of SMC3, STAG1 and STAG2 immunoprecipitates obtained from chromatin extracts of cells synchronized in G1 and G2. Figure 1 continued on next page analyzed without SMC3 normalization. Even though under these conditions more cohesin STAG2 was present than cohesin STAG1 , less acetylated SMC3 was detected in cohesin STAG2 than in cohesin STAG1 ( Figure 1-figure supplement 1C). In contrast, cohesin STAG1 and cohesin STAG2 contained similar amounts of acetylated SMC3 in G2 phase ( Figure 1A; compare lanes 5 and 7). These results indicate that SMC3 acetylation in G1 occurs preferentially on cohesin STAG1 .
Supporting this hypothesis, we observed that STAG1 depletion by RNAi reduced SMC3 acetylation in G1 to a similar extent as depletion of ESCO1, whereas STAG2 depletion had little effect ( Figure 1B). This experiment also revealed that more STAG1 accumulates on chromatin in STAG2depleted cells than in control-depleted cells ( Figure 1B, lane 4). We also observed elevated levels of STAG1 in whole cell lysates prepared from STAG2-depleted cells (Figure 1-figure supplement 1D, compare lanes 1 and 10), indicating that cells compensate for loss of STAG2 by increasing STAG1 levels by an unknown mechanism. However, this compensation is only partial as SCC1 levels were approximately three-fold lower in STAG2 depleted cells than in control cells (Figure 1-figure supplement 1D; compare lanes 1 and 10). This difference is important for the interpretation of Hi-C results which will be described below.
To determine the acetylation levels of cohesin STAG1 and cohesin STAG2 , we used label-free quantitative mass spectrometry (qMS). To be able to isolate cohesin STAG1 and cohesin STAG2 under comparable conditions we generated HeLa cells in which all STAG1 or STAG2 alleles were modified using CRISPR-Cas9-mediated genome editing to encode enhanced green fluorescent (EGFP) fusion proteins (for characterization of these cell lines, see Figure 2-figure supplement 1A) and isolated cohesin STAG1 and cohesin STAG2 using antibodies to GFP ( Figure 1C). This analysis indicated that in G1 cohesin STAG1 contains four times more acetylated SMC3 than cohesin STAG2 ( Figure 1D).
A subpopulation of cohesin STAG1 associates stably with chromatin during G1 phase Because SMC3 acetylation stabilizes cohesin on chromatin in S and G2 phase (Ladurner et al., 2016), we tested whether acetylated cohesin STAG1 also has a longer residence time on chromatin in G1 than the less acetylated cohesin STAG2 complexes. To analyze the chromatin binding dynamics of cohesin STAG1 , we performed inverse fluorescence recovery after photobleaching (iFRAP) in STAG2depleted G1 cells that expressed a GFP-tagged version of SMC3 (SMC3-LAP; for depletion efficiency, see Figure 2-figure supplement 1B). Since STAG proteins are required for cohesin's association with chromatin (Roig et al., 2014), the behavior of SMC3-LAP in these cells should predominantly reflect the behavior of cohesin STAG1 . Conversely, to analyze cohesin STAG2 we analyzed SMC3-LAP in cells depleted of STAG1. These experiments confirmed previous observations (Gerlich et al., 2006) that in control cells most cohesin interacts with chromatin dynamically with a residence of 13 min (Figure 2A,B and E). Similar results were obtained in cells depleted of STAG1, implying that in G1 most cohesin STAG2 interacts with chromatin dynamically (Figure 2A,B and E). In contrast, in cells depleted of STAG2, the SMC3-LAP signal equilibrated much more slowly between bleached and unbleached regions (Figure 2A and B) and the resulting iFRAP curve could only be Immunoprecipitated material was normalized to SMC3 levels and immunoblotting was performed using the indicated antibodies. Flow cytometry profiles are shown on the left. (B) Flow cytometry and chromatin extract immunoblot analysis of cells synchronized in G1 and depleted of the proteins indicated. Chromatin extracts were normalized relative to SMC3. Note that SMC3ac levels are decreased after ESCO1 and STAG1 depletion and increased after STAG2 depletion. (C) SDS-PAGE analysis of EGFP-STAG1 and STAG2-EGFP immunoprecipitates obtained from chromatin extracts of cells synchronized in G1 and G2. Immunoprecipitations were performed using anti-GFP antibodies and were analyzed with silver staining. Flow cytometry profiles are shown on the left. (D) Relative abundance of non-acetylated peptides (no acetylation), peptides acetylated at position K105 (K105ac) and peptides acetylated peptide at positions K105 and K106 (K105ac/106ac) in cells synchronized in G1 and G2 was determined by quantitative mass spectrometry from material immunoprecipitated using anti-GFP antibodies from EGFP-STAG1 and STAG2-EGFP cells. Peptide percentage was calculated relative to the total number of SMC3 peptides. Note that part of this experiment has been previously published (Ivanov et al., 2018; Figure S4). The online version of this article includes the following figure supplement(s) for figure 1:        fitted to a biexponential function (Figure 2-figure supplement 1C). This indicates that cohesin-STAG1 complexes exist in two distinct populations which interact with chromatin differently. Analysis of these data revealed that most cohesin STAG1 (63%) is bound to chromatin dynamically with a residence time of 7 min, but that a smaller subpopulation of cohesin STAG1 (37%) is stably associated with chromatin with a residence time of 3 hr ( Figure 2C and D). The latter residence time is similar to that observed for cohesin after depletion of WAPL (Figure 2A-D; Tedeschi et al., 2013), suggesting that the stably chromatin bound cohesin STAG1 complexes are protected from release by WAPL.
To analyze the stability of cohesin STAG1 and cohesin STAG2 on chromatin in G1 more directly, we performed iFRAP experiments in CRISPR-generated EGFP-STAG1 and STAG2-EGFP knock-in cell lines ( Figure 2G-L, Figure 2-figure supplement 1A and D). As predicted from our STAG2 depletion experiments in SMC3-LAP cells, the EGFP-STAG1 iFRAP curve could only be fitted to a biexponential function. A subpopulation (33%) of EGFP-STAG1 was bound to chromatin stably ( Figure 2I) with a residence time of 5 hr ( Figure 2J). The remaining 67% of EGFP-STAG1 bound dynamically to chromatin with a residence time of 15 min ( Figure 2K). In contrast, the STAG2-EGFP iFRAP curve could be fitted to a single exponential function, with a dynamic residence time of five minutes ( Figure 2L). A stably bound fraction of EGFP-STAG1 but not of STAG2-EGFP could also be detected in FRAP experiments, in which, compared to iFRAP, much smaller nuclear volumes are photobleached and therefore fluorescence recovery occurs on shorter time scales, resulting in higher temporal resolution (Figure 2-figure supplement 2). Importantly, these differences between STAG1 and STAG2 were not caused by N-terminal tagging of STAG1 since we obtained similar results in an independent cell line, in which STAG1 was C-terminally tagged with EGFP ( Figure 2-figure supplement 2; for technical reasons we were able to generate this cell line only late during this study, which is why earlier experiments were performed with a version of STAG1 which is tagged on its N-terminus, that is, different to the C-terminally tagged STAG2). These results indicate that a small subpopulation of cohesin STAG1 stably associates with chromatin in G1 in HeLa cells. We suspect that previous studies (Gerlich et al., 2006) failed to detect these complexes because they only represent about 9% of all cohesin complexes (37% of cohesin STAG1 ; Figure 2C, which represents 25% of all cohesin; Holzmann et al., 2019).
Since this stable population of G1 cohesin has not previously been described, we tested whether such a population also exists in mammalian cells other than HeLa. We therefore performed Stag2 RNAi in immortalized mouse embryonic fibroblasts (iMEFs) and monitored the fluorescence recovery of Scc1-LAP ( Key resource table). In these cells, we were able to detect stably chromatin bound cohesin in G1 even without Stag2 depletion, perhaps because in iMEFs cohesin Stag1 represents 33% of total cohesin (Remeseiro et al., 2012). We observed that approximately 20% of Scc1-LAP and 12% of Smc1-LAP were stably bound to chromatin in G1 in these cells ( The stable association of cohesin STAG1 with chromatin depends on ESCO1 and CTCF, as does SMC3 acetylation Since stabilization of cohesin in S and G2 depends on ESCO proteins and sororin (Ladurner et al., 2016;Schmitz et al., 2007), we tested whether the same proteins are also required for stabilization of cohesin STAG1 in G1 (note that even though sororin levels are very low in G1, some sororin can be detected during this cell cycle phase; Figure S2 in Nishiyama et al., 2010). To this end, we depleted sororin or ESCO1 by RNAi in G1 ( Figure 3A) and measured the recovery of EGFP-STAG1 using iFRAP. Sororin depletion did not affect the chromatin binding dynamics of cohesin STAG1 ( Figure 3B and C), as one might have expected given the low levels of sororin in G1 (Nishiyama et al., 2010;Rankin et al., 2005). This was not due to insufficient depletion of sororin because the same siRNA oligomer reduced the stable binding of EGFP-STAG1 to chromatin in G2 (Figure 3-figure supplement 1A-E). In contrast, depletion of ESCO1 converted most stably bound cohesin STAG1 complexes into dynamic ones ( Figure 3B and C). This result was also observed using a second siRNA oligomer     , suggesting that acetylation of cohesin STAG1 complexes is required for their stable association with chromatin. Unexpectedly, while characterizing cohesin STAG1 and cohesin STAG2 immunoprecipitates, we noticed that CTCF was more abundant in cohesin STAG1 samples, as detected by immunoblotting ( Figure 3D) and label-free quantitative mass spectrometry (Figure 3-figure supplement 2). We therefore tested whether CTCF is also required for stable binding of cohesin STAG1 to chromatin. To our surprise, CTCF depletion indeed reduced stable chromatin binding of EGFP-STAG1 to a degree like that observed following ESCO1 depletion ( Figure 3F and G). Co-depletion of WAPL with CTCF reverted this effect ( Figure 3H-J), indicating that CTCF contributes to long chromatin residence times of cohesin STAG1 by protecting it from WAPL. Similar results were obtained by performing FRAP in a cell line in which STAG1 was tagged with EGFP and STAG2 was tagged with red fluorescence protein (RFP) at their endogenous loci ( Interestingly, we found that depletion of CTCF also strongly reduced SMC3 acetylation levels ( Figure 3E), as we had previously observed in primary Ctcf 'knockout' MEFs arrested in G0; see Figure 1A in Busslinger et al. (2017). This indicates that CTCF is also required for cohesin acetylation in G0 and G1.

Acetylated cohesin is enriched at loop anchors
Cohesin's residence time on chromatin is thought to determine the lifetime and distance of chromatin interactions (Fudenberg et al., 2016;Gassler et al., 2017;Haarhuis et al., 2017;Wutz et al., 2017). Because some cohesin STAG1 complexes have long chromatin residence times, we analyzed whether cohesin STAG1 complexes contribute to chromatin architecture differently during G1 than cohesin STAG2 complexes, which have short residence times.
For this purpose, we first analyzed in ChIP-seq experiments where in the genome cohesin STAG1 is enriched compared to cohesin STAG2 . To avoid artefacts caused by using different antibodies, we used GFP antibodies in cell lines in which endogenous STAG1 and STAG2 had been tagged with EGFP ( Figure 2-figure supplement 1A). This revealed that most STAG1 and STAG2 peaks the fraction of nuclear EGFP-STAG1 that was stably chromatin bound in cells synchronized in G1 and depleted of the indicated proteins by RNAi. Quantification of dynamic and stable residence time of EGFP-STAG1 is shown on the right. (H) Immunoblot analysis of whole cell extract from cells depleted of CTCF, WAPL or CTCF/WAPL double depletion. a-tubulin was used as a loading control. (I) Graph depicting the mean normalized difference in EGFP-STAG1 fluorescence intensity between the unbleached and bleached regions following iFRAP in G1 cells and depleted of the indicated proteins by RNAi. Error bars denote s.e.m., n = 10 cells per condition. (J) Quantification of the fraction of nuclear EGFP-STAG1 that was stably chromatin bound in cells synchronized in G1 and depleted of the indicated proteins by RNAi. Quantification of dynamic and stable residence time of EGFP-STAG1 upon RNAi treatment for indicated proteins is shown on the right. Note that analyses of the data in I yielded lower values for stable residence times than those in C and G, presumably as a result of the shorter iFRAP imaging time used in this experiment. The online version of this article includes the following source data and figure supplement(s) for figure 3: Source data 1. The Microsoft Excel file lists iFRAP measurements used to generate data in Figure 3B,C,F,G. Source data 2. The Microsoft Excel file lists iFRAP measurements used to generate data in Figure 3I-J.      Mb region of HeLa chromosome 1. ChIP-seq signals for EGFP-STAG1, STAG2-EGFP, SMC3, SMC3ac and CTCF are shown below the Hi-C contact matrix. STAG1 and STAG2 immunoprecipitation was performed using anti-GFP antibodies. Note that the SMC3ac signal is decreased following ESCO1 depletion, confirming the specificity of the anti-SMC3ac antibody. (B) Venn diagram illustrating the genome-wide co-localization between STAG1 and STAG2 ChIP-seq signals. STAG1/2 immunoprecipitations from EGFP-STAG1 and STAG2-EGFP cells were performed using anti-GFP antibodies. Average read density plots for EGFP-STAG1 and STAG2-EGFP at SMC3ac and SMC3 sites is shown on the right. overlapped (74% of all peaks; Figure 4A and B, left panel). As recently reported for mouse cells (Kojic et al., 2018), sites at which predominantly STAG2 was found overlapped more frequently with transcription start sites (TSSs; 11.9%) and enhancers (5.5%) than STAG1-only sites, of which only 2% and 0.05% overlapped with TSSs and enhancers, respectively ( Figure 4C). Kojic et al. (2018) also reported that STAG1-only sites overlap with CTCF sites more frequently than STAG2-only sites.
Although we observed a similar tendency, we found that most STAG1-only and STAG2-only sites did not overlap with CTCF, whereas most common sites did ( Figure 4D). Next, we determined where in the genome stably chromatin bound cohesin STAG1 complexes are enriched. To address this, we took advantage of our finding that many if not all of these complexes are acetylated on SMC3 (Figure 1). We could detect acetyl-SMC3 at 24% of all SMC3 peaks (hereafter called pan-SMC3), at 39% of common STAG1 and STAG2 peaks and at 27% of CTCF peaks. Conversely, practically all SMC3-ac overlapped with SMC3, common STAG1 and STAG2 sites and CTCF ( Figure 4E) but did not overlap with STAG1-only or STAG2-only sites ( Figure 4F). At the common STAG1 and STAG2 sites at which acetyl-SMC3 could be detected, the read density of STAG1 was higher than that of STAG2 ( Figure 4B, right panel), consistent with a longer residence time and thus higher enrichment of some of the cohesin STAG1 complexes. Importantly, the number of acetyl-SMC3 peaks was reduced by ESCO1 depletion from 15,229 to 8,850, indicating that this antibody preferentially recognizes the acetylated form of SMC3 and not just unmodified SMC3 with reduced affinity (for a representative example, see Figure 4A).
We then used Hi-C to generate high resolution genome architecture maps in wild type HeLa cells (Hi-C map 1 in Supplementary file 1; 916 million unique read pairs). By comparing the ChIP-seq profiles of pan-SMC3 and acetyl-SMC3 with these maps, we observed that acetyl-SMC3 was more frequently found at loop anchors (74% of SMC3ac overlapped with loop anchors) than pan-SMC3 (43%; Figure 4A and G). We suspect that this difference is an underestimate of the specific enrichment of acetyl-SMC3 at loop anchors, as the pan-SMC3 profile presumably represents the sum of unmodified and acetylated SMC3, and because due to the resolution of our Hi-C maps, the loop anchors were on average 9 kb long and contained on average three SMC3 peaks and one acetyl-SMC3 peak. In other words, an even higher resolution Hi-C map would be expected to reveal an even stronger enrichment of acetyl-SMC3 at loop anchors.

Cohesin STAG1 generates long chromatin interactions
Our data obtained so far indicated that, in G1, ESCO1 predominantly acetylates cohesin STAG1 , that some of these complexes bind to chromatin stably in an ESCO1 dependent manner, and that acetylated cohesin is enriched at loop anchors. We therefore hypothesized that cohesin STAG1 might be particularly important for forming long chromatin interactions.
To test this possibility, we depleted STAG1 or STAG2 by RNAi in cells synchronized in G1 (Figure 5-figure supplement 1A) and generated high-resolution Hi-C maps (Hi-C maps 2 and 3 in Supplementary file 1, >840 million unique pairs each). Following STAG1 depletion only minor changes could be observed, but following STAG2 depletion, that is in the presence of cohesin STAG1 , new loops became detectable ( Figure 5A). This is remarkable given that our analysis of SCC1 had indicated that these cells contained three-fold less cohesin (see Figure 1-figure supplement 1D above). When we plotted the cumulative proportion of 7177 loops only detected in STAG2-depleted cells as a function of their length, we found that they were longer than control-specific loops of which we identified 14726 ( Figure 5B), confirming that cohesin STAG1 complexes form longer loops than cohesin STAG2 .
Corresponding changes could also be seen in contact probability plots ( Figure 5-figure supplement 1B). Depletion of either STAG1 or STAG2 reduced the contact probability over short genomic distances (10-100 kb) but increased them over longer distances (>100 kb) compared to control cells, with STAG2 depletion having a much stronger effect. In control cells, the genomic distance with highest contact probability was around 200 kb, corresponding to the bulk of interactions that are To exclude the possibility that RNAi artefacts such as off target effects could have contributed to these phenotypes, we also generated cell lines in which either STAG1 or STAG2 could be inactivated by auxin inducible degradation (AID; Figure 5-figure supplement 1E; note that AID tagging reduced STAG1 and STAG2 levels already in the absence of auxin, possibly due to 'leakiness' of the AID system, implying that differences between Hi-C phenotypes obtained with and without auxin treatment might be smaller than they would be otherwise). Also, in this case Hi-C experiments revealed only minor changes after STAG1 degradation, but following STAG2 depletion, that is in the presence of cohesin STAG1 , new loops became detectable (Hi-C maps 15-18 in Supplementary file 1, >250 million unique pairs each; Figure 5-figure supplement 1F). These observations indicate that cohesin STAG1 mediates the formation of chromatin interactions that are longer than the ones formed by cohesin STAG2 .
Fluorescence in situ hybridization experiments support the population Hi-C data obtained from STAG1-depleted and STAG2-depleted cells  Figure 5-figure supplement 2C for representative images from each experimental condition). We also performed the same analysis with a probe pair not predicted to span a loop ( Figure 5-figure supplement 2A, last example). We used automated image analysis to measure the three-dimensional distance between each pair of probes in more than 100 cells in each experimental condition (number of cells analyzed per condition and test of statistical significance between control and different conditions are listed in Supplementary file 2). This inter-probe distance was variable for each pair of probes, either reflecting technical variability, and/or indicating that the length of each predicted loop differed between cells, consistent with previous FISH and single-cell Hi-C studies Nagano et al., 2013;Nora et al., 2017). In all six test loops, but not in the control genomic region, the inter-probe distance increased following depletion of SCC1 or following depletion of STAG1 in combination with STAG2. Depletion of CTCF also led to an increase in inter-probe distance in five out of six test loops, consistent with CTCF's proposed function as a boundary for loop formation. This indicates that this experimental setup can detect cohesin and CTCF-specific changes in chromatin architecture in single cells. Importantly, in five out of six test loops, depletion of STAG1 alone led to a greater increase in inter-probe distance than depletion of STAG2. This is consistent with our were down-sampled to 314 million reads). Matrices were plotted using Juicebox. (D) Coverage-corrected Hi-C contact matrices of chromosome 10 (60-62 Mb), in control-, STAG1-, STAG2-and WAPL-depleted HeLa cells (libraries with numbers 7, 10, 11 in Supplementary file 1; library from the WAPL RNAi was previously published in Wutz et al., 2017). (E) The proportion of convergent, tandem and divergent CTCF binding orientation for loops with both anchors overlapping SMC3 and CTCF ChIP-seq peaks as well as unambiguous CTCF binding directions, for loops identified in control-depleted cells in G1-phase (ctrl, 2799 loops), in WAPL depleted but not in control-depleted cells in G1-phase (WAPL, 2295 loops), in STAG1 depleted but not in control-depleted cells in G1-phase (STAG1, 367 loops) and STAG2 depleted but not in control-depleted cells in G1-phase (STAG2, 1471 loops). The theoretically expected random proportions assuming no directionality bias are shown as comparison (7,10 and 11 in Supplementary file 1; library for the WAPL RNAi was previously published in Wutz et al., 2017). (F) Intra-chromosomal contact frequency distribution as a function of genomic distance using logarithmically increasing genomic distance bins for control-, STAG1-, STAG2-and WAPL-depleted HeLa cells. (G) Distribution of loop length in control-, STAG1-, STAG2-and WAPL-depleted HeLa cells. Loops identified by hiccups-(Hi-C libraries 7,10 and 11 in Supplementary file 1; library for the WAPL RNAi was previously published in Wutz et al., 2017). The online version of this article includes the following figure supplement(s) for figure 5:  hypothesis that cohesin STAG1 is more important for generating longer-range chromatin loops than cohesin STAG2 .

Long-range chromatin interactions mediated by cohesin STAG1 are similar but not identical to those observed in WAPL depleted cells
Many of the loops that could only be detected in STAG2-depleted cells, that is were presumably mediated by cohesin STAG1 , had loop anchors in the outer boundaries of two or more adjacent TADs ( Figure 5A), reminiscent of long-range chromatin interactions observed in WAPL depleted cells (Haarhuis et al., 2017;Wutz et al., 2017). We therefore compared Hi-C interactions in STAG2depleted and WAPL-depleted cells.
For this purpose, we performed two different comparisons. In one case, we generated a new medium-resolution Hi-C map from WAPL depleted cells (Hi-C map four in Supplementary file 1) and compared this to the high-resolution Hi-C maps from control, STAG1 and STAG2 depleted cells (Hi-C maps 1-3 in Supplementary file 1; for comparability, unique sequence reads in the three latter data sets were randomly downsampled to the number of reads obtained from WAPL depleted cells; Figure 5C and Figure 5-figure supplement 1E and F). In the other case, we generated low-resolution libraries from control, STAG1-depleted and STAG2-depleted cells in replicate, using the same experimental conditions and protocols as those previously used to study the consequences of WAPL depletion ( Figure 5D; Hi-C maps 7, 10 and 11 in Supplementary file 1, around 480 million read pairs each) and compared these to the previously generated replicate Hi-C maps from WAPLdepleted cells (Wutz et al., 2017). These comparisons revealed interesting similarities and differences between STAG2-depleted and WAPL-depleted cells.
First, we observed that most loops specifically detected in STAG2-depleted cells still 'obeyed' the CTCF convergence rule ( Figure 5E and Figure 5-figure supplement 1G), whereas depletion of WAPL led to partial violation of this rule ( Figure 5E and Figure 5-figure supplement 1G), as previously reported (Haarhuis et al., 2017;Wutz et al., 2017). These results indicate that cohesin STAG1 forms long-range chromatin interactions that are anchored at convergent CTCF sites and raises the interesting possibility that WAPL is required for the CTCF convergence rule (see Discussion). Second, the genomic distance with highest contact probability was similar in STAG2-depleted cells (800 kb) and in WAPL depleted cells (900 kb), i.e., much longer than in control cells (200 kb) and in STAG1 depleted cells (300 kb; Figure 5F). Third, more loops > 500 kb could be detected in STAG2depleted cells by hiccups than in control cells, similar to the situation in WAPL depleted cells where an even higher number of loops > 500 kb could be detected. In contrast, the number of loops > 500 kb was reduced in STAG1-depleted cells ( Figure 5G and Figure 5-figure supplement 1H).
Together, these observations support the notion that acetylated cohesin STAG1 complexes are protected from WAPL, therefore have longer residence times on chromatin and can form longer chromatin interactions, which, however, still obey the CTCF convergence rule.

In silico modeling indicates that long chromatin residence time of cohesin STAG1 causes formation of long chromatin loops
To test the hypothesis that an increased chromatin residence time of cohesin STAG1 causes the formation of long chromatin loops we performed simulations in silico. For this purpose, we used a simplified hypothetical DNA sequence that contained three pairs of convergent CTCF sites and assumed three different extrusion times ( Figure 6A). We also used molecular dynamics simulations to model the behavior of chromatin in the presence of loop extrusion complexes and generated an insilico contact map of a region of human chromosome 9. This map resembled the Hi-C map of this region generated from control HeLa cells ( Figure 6B and C, compare top and bottom panels). We next simulated the effect of altering the lifetime of loop extrusion complexes on DNA and found, consistent with results by Fudenberg et al. (2016) , that longer lifetimes resulted in the generation of longerrange interactions, and vice versa. These in silico contact maps reproduced the changes we observed in cells following STAG1 and STAG2 depletion reasonably well. Thus, our experimental and simulation data are consistent with the hypothesis that cohesin STAG1 and cohesin STAG2 contribute to chromatin organization differentially, and that, by virtue of its longer residence time on chromatin, cohesin STAG1 generates longer-range interactions than cohesin STAG2 .

ESCO1, like CTCF, regulates the loop formation activity of cohesin STAG1
The hypothesis that stably chromatin bound cohesin STAG1 complexes form long chromatin loops predicts that also ESCO1 and CTCF are required for these interactions, since we had found that both are required for the long residence time of cohesin STAG1 ( Figure 3A-G). To test this prediction, we co-depleted ESCO1 or CTCF together with STAG2, synchronized cells in G1 and performed Hi-C analysis (Figure 7, Figure 7-figure supplement 1; Hi-C maps 7, 8, 9, 11, 12 and 13 in Supplementary file 1, around 200 million unique read pairs each). Indeed, we found that the appearance of new loops in STAG2 depleted cells was largely reverted by co-depletion of ESCO1 or CTCF (Figure 7-figure supplement 1C-F). This epistatic behavior of ESCO1 and CTCF depletion over STAG2 depletion supports the hypothesis that the long residence time of cohesin STAG1 enables the formation of long chromatin loops in G1. We also analyzed cells from which only CTCF or ESCO1 had been depleted (Figure 7). Consistent with previous results obtained by AID of CTCF (Nora et al., 2017;Wutz et al., 2017), depletion of CTCF by RNAi resulted in a reduction in the number of detectable loops ( Figure 7A,C and D), and a decrease in the TAD insulation score ( Figure 7E) but did not abolish long-range chromatin interactions. To the contrary, contact probability analysis revealed that the interactions involved in TAD formation were longer in cells depleted of CTCF ( Figure 7F; note that this particular effect was less pronounced after auxin induced CTCF degradation, perhaps because in these experiments CTCF levels were also reduced in control cells due to 'leakiness' of the AID system; Wutz et al., 2017). As concluded previously (Nora et al., 2017;Wutz et al., 2017), these results suggest that CTCF is not required for long-range chromatin interactions per se but for specifying the loop anchors which mediate these, possibly by functioning as a boundary for loop extruding cohesin complexes. Remarkably, ESCO1 depletion caused similar effects, that is a reduction in the number of detectable loops ( Figure 7A,C and D), a decrease in the TAD insulation score ( Figure 7E) and an increase in the length of chromatin interactions ( Figure 7F). This raises the interesting possibility that ESCO1, like CTCF, is important for restricting cohesin's loop formation activity (see Discussion).
However, a comparison of Hi-C phenotypes between ESCO1-depleted and CTCF-depleted cells did not only reveal similarities but also differences. First, compared to both CTCF-depleted and control cells, contact frequencies were reduced in the 0-50 kb range following ESCO1 depletion ( Figure 7F). Second, ESCO1 depletion also resulted in an increase in contact probability around 20 Mb, the genomic distance associated with compartmentalization ( Figure 7F). Consistently, the analysis of whole chromosome Hi-C maps revealed that depletion of ESCO1, either alone or in combination with STAG2 depletion, enhanced the 'checkerboard' pattern indicative of compartmentalization and led to a strong increase in interactions around 10 Mb ( Figure 7B). Genome-wide aggregate analysis of 50 compartment categories ranging from strong B to strong A compartments confirmed this, showing increasing contact enrichment between similar compartment categories and a decreasing contact enrichment between dissimilar (e.g., strong A and strong B) compartment bins in both long cis (>2 Mb) and trans interactions (Figure 7-figure supplement 1G). This phenotype is reminiscent of the increase in compartmentalization observed following cohesin depletion Rao et al., 2017;Schwarzer et al., 2017;Wutz et al., 2017), even though ESCO1-depleted cells contained as much cohesin on chromatin as the corresponding control of STAG2-depleted cells (Figure 7-figure supplement 1A, compare SCC1 signals in lanes 3 and 6, and in lanes 9 and 12). These observations raise the interesting possibilities that the ability of cohesin to suppress compartmentalization depends on ESCO1, or that this acetyltransferase has additional functions in chromatin organization that are independent of cohesin.

Discussion
Acetylated cohesin STAG1 complexes are protected from WAPL by CTCF and form long chromatin loops In interphase cells, cohesin folds genomic DNA into thousands of loops which are thought to have both structural and regulatory functions. Little is known about the lifetime of these loops and how their formation and maintenance is controlled. It has generally been assumed that loops are shortlived dynamic structures because the cohesin complexes that form them interact with DNA only briefly, in mammalian cells on average for 8-25 min during G0 and G1 phase (Gerlich et al., 2006; Tedeschi et al., 2013), after which they are released by WAPL (Kueng et al., 2006). Despite this, long-range chromatin interactions as they can be detected by Hi-C change little over time (Nagano et al., 2017;Wutz et al., 2017), and recent evidence implies that some loops can persist for hours (Vian et al., 2018). It is not known whether these stable structures are maintained by dynamically exchanging cohesin complexes or by an unknown mechanism that would protect cohesin from release by WAPL. Precedence for the latter scenario comes from the observation that cohesin complexes that mediate cohesion in proliferating cells are protected from WAPL by acetylation of their SMC3 subunit (Rolef Ben-Shahar et al., 2008;Unal et al., 2008), by subsequent recruitment of sororin and at mitotic centromeres also by shugoshin (Hara et al., 2014;Nishiyama et al., 2010). However, these cohesive complexes do not seem to participate in loop formation, as their stabilization on chromatin does not detectably alter chromatin structure in G2 phase (Wutz et al., 2017), whereas experimental stabilization of all cohesin complexes on chromatin does (Tedeschi et al., 2013;discussed in Holzmann et al. (2019). It has therefore remained unknown whether loop forming cohesin complexes can be protected from WAPL to extend the lifetime of chromatin loops.
Here we provide evidence that such a regulatory mechanism exists in human cells, since our FRAP and Hi-C experiments have identified a small subpopulation of cohesin STAG1 complexes that persist on chromatin for hours and contribute to the formation of long chromatin loops. Our results indicate that the stabilization of these cohesin STAG1 complexes on chromatin depends on SMC3 acetylation, as does the stabilization of cohesive cohesin in S and G2 (Ladurner et al., 2016). But in contrast to cohesive cohesin, loop forming cohesin STAG1 complexes can persist on chromatin for hours in the absence of sororin, as one might have predicted since sororin is only present in proliferating cells, and in these almost exclusively from S phase until mitosis (Nishiyama et al., 2010;Rankin et al., 2005). Likewise, we suspect that shugoshin is dispensable for the long chromatin residence time of acetylated cohesin STAG1 , since shugoshin specifically protects cohesive cohesin from WAPL at centromeres in mitosis (Hara et al., 2014). Unexpectedly, however, we found that CTCF is essential for the long residence time of cohesin STAG1 on chromatin, indicating that CTCF, like sororin and shugoshin, is a WAPL antagonist that can prevent cohesin from being released from DNA ( Figure 8A).
The notion that cohesin STAG1 complexes are protected from WAPL by SMC3 acetylation and CTCF is supported by the observation that the chromatin residence time of these complexes is similarly long as the residence time of cohesin in WAPL depleted cells (Figure 2), by our finding that cohesin STAG1 complexes form longer chromatin loops as cohesin does in the absence of WAPL (Figure 5), and by the epistatic effect of WAPL depletion over CTCF depletion ( Figure 3H). Interestingly, however, the extended loops formed by cohesin STAG1 differ in one important aspect from the loops that are formed by cohesin in the absence of WAPL, in that the former are typically anchored at convergent CTCF sites ( Figure 5E and Figure 5-figure supplement 1G) whereas the latter are often anchored at tandemly oriented CTCF sites (Haarhuis et al., 2017;Wutz et al., 2017; Figure 5E and Figure 5-figure supplement 1G). This difference implies that WAPL may contribute to the CTCF convergence rule, that is may have a role in ensuring that loops are only anchored at convergent CTCF sites.
Our photobleaching experiments also revealed that most if not all cohesin STAG2 complexes, which are three fold more abundant in HeLa cells than cohesin STAG1 complexes (Holzmann et al., 2019), have short chromatin residence times in the range of minutes, implying that they are dynamically released from chromatin by WAPL. Accordingly, CTCF depletion only had a small effect on their chromatin residence time. However, this does not exclude the possibility that CTCF would also be able to protect cohesin STAG2 from WAPL under conditions where these complexes become stabilized on chromatin. The existence of two different forms of cohesin with different chromatin residence times is reminiscent of the situation in mitotic chromosomes. Their structural organization is thought to depend on the sequential action of first condensin II to form large loops and then condensin I to form smaller, nested loops (Gibcus et al., 2018;Walther et al., 2018). It is unknown how interphase chromatin architecture is established upon exit from mitosis, but recent fluorescence correlation spectroscopy experiments indicate that STAG1 is recruited to chromatin earlier than STAG2 (Cai et al., 2018). Given our finding that cohesin STAG1 generates longer loops than cohesin STAG2 , it is tempting to speculate that cohesin STAG1 and cohesin STAG2 function analogously to condensin II and condensin I in mitosis, respectively, by forming loops within loops on interphase chromatin ( Figure 8B).
Consistent with our findings, it has recently been reported that cohesin STAG1 and cohesin STAG2 contribute to chromatin differently (Casa et al., 2019;Viny et al., 2019;Cuadrado et al., 2019;Kojic et al., 2018), with cohesin STAG1 contributing more to TAD organization than cohesin STAG2 . Cohesin STAG1 has also been shown to be more resistant to biochemical salt extraction from chromatin than cohesin STAG2 (Kojic et al., 2018), but whether this property is related to the long chromatin residence time of acetylated cohesin STAG1 reported here is unknown. Both our study and that of Kojic et al. (2018) provide evidence that STAG1 and STAG2 are enriched at CTCF sites and enhancers, respectively. This raises the possibility that cohesin STAG2 might regulate promoterenhancer interactions. STAG2 is one of only twelve genes known to be mutated in more than four major human cancer types (Lawrence et al., 2014). It is therefore possible that changes in gene expression following STAG2 mutation might be a common early event in human carcinogenesis.

ESCO1 is needed to constrict cohesin at CTCF sites
The acetylation that protects cohesive cohesin from WAPL from S phase until mitosis is thought to depend on both ESCO1 and ESCO2 (Ladurner et al., 2016;Nishiyama et al., 2010). In contrast, acetylation of cohesin during G1 is only mediated by ESCO1, since ESCO2 is absent during this phase of the cell cycle (Alomer et al., 2017;Minamino et al., 2015;Rahman et al., 2015;Whelan et al., 2012; Figure 1-figure supplement 1A). Our finding that ESCO1 is needed for a long chromatin residence time of a subpopulation of cohesin complexes and for their ability to form extended chromatin loops reveals for the first time a function of SMC3 acetylation in G0 and G1, and at the same time indicates that ESCO1 has a certain degree of substrate specificity, at least during G1 phase, by preferentially acetylating cohesin STAG1 .
Interestingly, our results revealed that ESCO1 might also have another function in chromatin organization. ESCO1 depletion reduced the Hi-C dots and corner peaks specifically detected after STAG2 depletion, that is long chromatin loops formed by cohesin STAG1 , as one would have predicted given that ESCO1 is required for the long residence time of a subpopulation of cohesin STAG1 . However, in addition, ESCO1 also decreased TAD insulation, to an extent as seen after CTCF depletion. This implies that ESCO1 might have two functions in chromatin organization: protection of cohesin STAG1 from WAPL, and a boundary function in loop extrusion. If ESCO1 has such a boundary function, this could help to explain why co-depletion of ESCO1 and ESCO2 with WAPL resulted in the formation of vermicelli in chicken DT40 cells, whereas in these cells WAPL depletion alone did not cause this phenotype (Kawasumi et al., 2017). It is possible that in these experiments WAPL depletion increased the residence time of cohesin on chromatin, whereas depletion of ESCO1 compromised the function of CTCF boundaries, so that extruding cohesin complexes could form longer loops and ultimately accumulate in vermicelli domains.
The hypothesis that ESCO1 contributes to boundary function could also explain why PDS5A and PDS5B are required for boundary function (Wutz et al., 2017) because SMC3 acetylation by ESCO1 and ESCO2 has been shown to depend on PDS5A and PDS5B (Minamino et al., 2015; for a similar dependency in yeast see Chan et al., 2013 andVaur et al., 2012). It is therefore possible that one function of PDS5 proteins in chromatin architecture is to facilitate SMC3 acetylation by ESCO1. In addition, PDS5 proteins might influence chromatin architecture by other mechanisms as they are also required for WAPL-mediated release of cohesin from chromatin (Chan et al., 2012;Ouyang et al., 2016;Shintomi and Hirano, 2009;Wutz et al., 2017). Interestingly, we also observed a reduction in SMC3 acetylation following depletion of CTCF, indicating that SMC3 acetylation also depends on CTCF, and that the functions of ESCO1 and CTCF at chromatin boundaries might be interdependent (Busslinger et al., 2017;this study).
Consistent with our results, a role for the yeast ortholog of ESCO1 and ESCO2 (Eco1) in constraining cohesin's ability to form mid-range chromatin interactions, together with a dual role for the PDS5A and PDS5B ortholog Pds5 in regulating Eco1 and Wapl was reported during preparation of this manuscript (Dauban et al., 2020). The molecular mechanisms through which STAG1 and STAG2 alter the properties of cohesin remain to be understood, but our results suggest that cohesin STAG1 is more frequently stabilized at CTCF sites via PDS5A, PDS5B and ESCO1 than cohesin STAG2 . The regulation of chromatin boundaries by cohesin acetyltransferases and PDS5 proteins may therefore be an evolutionarily conserved mechanism, which might be spatially controlled in mammalian cells by CTCF.    Cell culture, cell synchronization and RNA interference HeLa Kyoto cells (Landry et al., 2013; RRID:CVCL_1922) were free from detectable mycoplasma contamination and have been authenticated by STR fingerprinting (Vienna Biocenter Core Facilities). Cells were cultured in DMEM supplemented with 10% FBS, 0.2 mM glutamine and penicillin/streptomycin (Gibco). Cells were synchronized at early S phase by two consecutive rounds of treatment with 2 mM thymidine (Sigma) and released into fresh media for 6 hr (G2) or 15 hr (G1). Synchronization was assessed by flow cytometry after methanol fixation and propidium iodide staining as described (Ladurner et al., 2014). Cells were treated with 30 nM siRNAs as indicated using RNAiMax (Invitrogen) at 48 or 72 hr before downstream analyses. Pre-annealed 21 nucleotide RNA with 3' double thymidine overhangs (Elbashir et al., 2001) was purchased from Ambion. Sense sequences for control, CTCF ('#1') (Wendt et al., 2008), ESCO1, ESCO2 (Nishiyama et al., 2010), Wapl ('Wapl1'), STAG1 ('SA1'), STAG2 ('SA2') (Kueng et al., 2006), and sororin (Schmitz et al., 2007) were denoted previously. ESCO1 second siRNA was a pool of 4 siRNAs: GGAAAGAGCAAACGAGG UA, GGACAGAAUAGCACGUAAA, CUAGAAGAGACGAAACGAA, GGACAAAGCUACAUGAUAG.

FRAP
For FRAP of SMC3-LAP, cells expressing fluorescent cohesin subunits were grown on chambered coverglass (Nunc 155409) for 1-3 days while treated with siRNAs and thymidine as indicated. Cells were imaged at 37˚C on a Zeiss LSM5 duo confocal microscope with 63x Plan-Apochromat objective and a 488 nm 100 mW diode laser for bleaching, or on a Zeiss LSM780 confocal microscope with 63x objective and bleaching with argon and diode lasers at 488 and 561 nm for dual color FRAP using CO2-independent media, or on an LSM880 confocal microscope (Carl Zeiss) with a 40 Â 1.4 NA oil DIC Plan-Apochromat objective (Zeiss) in cell culture medium without riboflavin and phenol red at 5% CO2. Cells were either cell cycle synchronized as described above, or G1 and G2 phase were identified by nuclear and cytoplasmic distribution of DHB-mKate2 signals, respectively.
Cycloheximide (1 ug/ml) was added before imaging to inhibit protein synthesis and contribution of new GFP expression to signal recovery. For spot FRAP, a circular region (r = 2 um) was bleached three times. Recovery of fluorescence was recorded over 10 min and 300 frames at 2 s intervals and normalized to 10 pre-bleach frames and background and cellular fluorescence measured with Fiji (Schindelin et al., 2012). Recovery curves were analyzed using Berkeley Madonna (www.berkeleymadonna.com). Curves were fitted by an exponential function with variables for free and transiently chromatin associated (Ladurner et al., 2014), dynamically and stably chromatin bound cohesin (Gerlich et al., 2006). Relative fractions and their residence times (reciprocal of the dissociation constant) were averaged and plotted using Prism software (GraphPad).
Inverse FRAP was used to specifically measure dynamics of cohesin bound to chromatin over several hours. To this end for SMC3-LAP (which is present in the nucleus and cytoplasm), an area covering the cell body was bleached except for a semicircle corresponding to approximately half of the nucleus, and cells were imaged intermittently using a motorized stage. Recovery was recorded over 2-4 hr at 3 min intervals and normalized as above. Curve fitting with single and bi-exponential function was used to deduce relative fractions and residence times of dynamic and stable cohesin on chromatin and plotted as above (see https://github.com/rladurner/STAG1/blob/master/curvefit. ipynb; Ladurner, 2020; copy archived at https://github.com/elifesciences-publications/STAG1). For STAG1-EGFP and EGFP-STAG2 (which show only nuclear GFP signal), iFRAP photobleaching was performed in half of nuclear regions with 2 iterations of 488 nm laser at max intensity after acquisition of two images. Fluorescence was measured in bleached-and unbleached regions followed by background subtraction with 1 min interval. iFRAP curves were normalized to the mean of the prebleach fluorescent intensity and to the first image after photobleaching. Curve fitting was performed with single exponential functions f(t)=EXP(-kOff1*t) or double exponential functions f(t)=a*EXP(-kOff1*t) +(1-a) *EXP(-kOff2*t) in R using the minpack.lm package (version 1.2.1). Dynamic and stable residence times were calculated from 1/kOff1 and 1/kOff2 respectively. Double exponential curve fitting was performed under constraint that 1/kOff1 and 1/kOff2 are in range between 1 min-40 min and 1.5 hr-15 hr respectively. Soluble fractions were estimated by the reduction of fluorescence signals in unbleached area after photobleaching.

Chromatin fractionation
Cells were extracted in a buffer consisting of 20 mM Tris-HCl (pH 7.5), 100 mM NaCl, 5 mM MgCl, 2 mM NaF, 10% glycerol, 0.2% NP40, 20 mM b-glycerophosphate, 0.5 mM DTT and protease inhibitor cocktail (Complete EDTA-free, Roche). Chromatin pellets and supernatant were separated and collected by centrifugation at 2,000 g for 5 min. The chromatin pellets were washed three times with the same buffer.

Immunoprecipitation
Cells synchronized in S phase by a thymidine treatment were released for 16 hr to G1 phase and harvested. Pellets were resuspended in lysis buffer and used for single-step IP of chromatin-bound fractions. The beads coupled to antibody and bound to responding proteins were mildly washed, and captured proteins were eluted with 0.1 M glycine, and eluates were neutralized with 1.5 M Tris-HCl pH9.2.

ChIP-seq
SMC3, GFP and CTCF ChIP-seq was performed as described in Wendt et al. (2008). SMC3(ac) ChIP was performed as in Schmidt et al. (2009). In brief, cell pellets were hypotonically treated, nuclei were isolated, lysed, and sonicated. Lysates were incubated overnight with protein-G dyna beads pre-bound to SMC3(ac) antibody (Nishiyama et al., 2010) or mouse non-immune IgG (SACSC-2025). Beads were washed, samples were eluted and de-crosslinked, DNA was purified and sequenced.

Mass spectrometry
To generate a peptide spanning lysin 105 and 106 the immunoprecipitates were digested in solution with 400 ng Glu-C (Sequencing Grade, Roche) at 37˚C for 16 hr.
The mass traces for the peptide VSLRRVIGAKKD in its unmodified, singly acetylated and doubly acetylated form were extracted from the raw files using the program Qualbrowser which is part of the Xcalibur software (Thermo Scientific). To account for different amounts of SMC3 protein between the immunoprecipitates the peptide area values were normalized based on the sum of the three most intense unmodified SMC3 peptides.

DNA FISH
DNA probes (BACs, Fosmids and Cosmids) were ordered at BAC Resources PAC and purified by midi-prep purification kit (Quiagen). Probes were labelled by nick translation using 1-2 mg DNA per 50 mL reaction (Sigma Aldrich). Probes were fluorescently labelled using Alexa dyes (Alexa-488, Alexa-568). Every 5 mL of nick-translated probe was ethanol precipitated together with 1 mL of salmon sperm DNA, 3 mL human Cot-1, 0.5 mL 3M sodium acetate and 60 mL ethanol 100%. Probes were then snap frozen in liquid nitrogen and centrifuged at 13000 rpm for 10 min at four degrees. The supernatant was carefully removed and replaced with 200 mL of ethanol 70%, and the probe mixture was centrifuged at 13000 rpm for 10 min at four degrees. The supernatant was carefully removed and the pellet air dried protected from light. The pellet was resuspended in 5 mL hybridization buffer (2xSSC, 20% w/v dextran sulfate, 50% formamide pH7) at 37 degrees for 10 min then denatured for 7 min at 80 degrees and incubated at 37 degrees for 30 min before use. HeLa cells were cultured, siRNA treated and synchronized on coverslips. Cells were quickly rinsed with PBS three times and fixed in a solution of PBS with 4% paraformaldehyde for 10 min at room temperature. Cells were then washed twice in PBS for 5 min each. Permeabilization of cells was performed in freshly made PBS, 0.5 Triton X-100 for 7 min on ice. Cells were washed twice with a solution of ethanol 70% for 5 min and dehydrated in 80%, 95% and 100% ethanol for 3 min each. Cells were then air dried and denatured in 50% formamide, 2xSSC adjusted at pH 7.2 for 30 min at 80 degrees. Cells were washed three times in cold 2xSSC. The coverslip was placed cell-side down onto the prepared fluorescently labelled probes on a slide and sealed with glue. Hybridization was performed overnight at 42 degrees in a dark and humid chamber. The glue was then removed carefully, and coverslips were placed cell-side up and washed protected from light three times in warm 50% formamide, 2xSSC pH 7.2 for 5 min each at 42 degrees and three times in 2xSSC for 5 min. Cells were briefly washed in 2xSSC at room temperature and counterstain in 0.2 mg/mL DAPI solution for 2 min at room temperature and washed twice in 2xSSC for 5 min each. Coverslips were dried and mounted in Vectashield and fixed with a minimal amount of nail varnish. Acquisitions were performed on LSM 880 and 780 confocal microscopes. After tile scan, hundreds of single nuclei positions were spotted using DAPI channel in X, Y and Z. Three-dimensional acquisitions were then made for each of these positions for all channels and saved for processing.
Batch Alleles Investigation Tool 'BAIT' was designed to run on Definiens software and measure FISH three-dimensional inter-probe distances automatically. Nucleus segmentation was performed in three-dimensions by use of the auto-threshold function of Definiens on DAPI staining. Nuclear boundaries were analyzed to detect and discard incomplete nuclei in X, Y and Z. Probes signals were determined as pixel intensity values within the nucleus. To comparison size and intensity of the signals, geometrical measurements were performed to set an object center 'Seed' and clusters of spots 'Allele' for and between each channel. Spot center distances were then calculated in three dimensions and exported as comma-separated values (.csv). Quality control was performed under Definiens to detect aberrations and artefacts. Statistics were performed using R. Violin distribution plots were made with Prism 8. Figures were made with Fiji standard deviation projection.

Comparison of FISH results with Hi-C maps
Numbers are given for the most densely populated square at highest resolution within the range of the loop coordinates targeted by our FISH probes. High-resolution Hi-C maps for control, STAG1-RNAi, and STAG2-RNAi at 5 Kb nominal resolution were generated using Juicebox Values indicate unique contact counts with KR (balanced) normalization, and no normalization applied.
Length distribution, CTCF occupancy, and CTCF orientation of loops The length of each loop was calculated as the distance between the midpoints of the loop anchors, and the length distributions of loops in each group were visualized with the empirical cumulative distribution function.
To determine CTCF occupancy, we expanded loop anchors smaller than 15 kb to 15 kb and counted the number of anchors that overlap with at least one CTCF ChIP-seq peak using bedtools. We calculated fold enrichment of CTCF occupancy by comparing the counts with the average overlap of ten random translational controls with the same length distribution as the loop anchors. Orientation of CTCF motifs at loop anchors was identified using MotifFinder in Juicer, and the proportion of inward oriented (following the convergent rule) CTCF motifs was reported.

Aggregate peak analysis (APA) of loops
We performed aggregate peak analysis of hiccups-called loops using juicer_tools apa. The aggregate enrichment of our sets of looping peaks in contact matrices was visualized by plotting a cumulative stack of sub-matrices around detected loop coordinates. For a map with 10 Kb resolution we generated squares of 210 Kb x 210 Kb summing all putative loop peaks in a way that the resulting APA plot displays the total number of contacts which lie within the entire loop set at the center of the matrix within the aggregate pile-ups of their surroundings (Rao et al., 2014). All underlying matrices are KR-normalized.

Aggregate analysis of TADs
For the analysis and visualization of average TAD pileups we generated size-sorted and -classified lists of TADs and calculated histogram matrices around the centers of those TAD areas within narrow predefined size ranges. That way several sub-matrices of interaction around similarly size-classified TADs were added up in order to generate a global profile of one size-range for every Hi-C matrix (10 Kb resolution, coverage-normalization).

Hi-C insulation plots
Plots of the insulation scores were made based on insulation bedGraph files and TAD boundary coordinates generated by using the 'findTADsAndLoops.pl' script in the HOMER software package. This software scans relative contact matrices for locally dense regions of contacts or areas with an increased degree of intra-domain interactions relative to surrounding regions. Using a resolution of 3000, a window size of 15000, and the default maximum interaction distance (2 MB), we generated a coordinate set of sites with maximal transition in contact orientation, that is sites with highest insulation. Average plots of insulation profiles in all samples were made from regions cantered around the coordinates in the respective bedGraph files.

Simulations
Molecular dynamic simulations were performed using HOOMD-blue (Anderson et al., 2008;Glaser et al., 2015) using an approach similar to Sanborn et al. (2015) with minor modifications. The region 17.65-19.25 mb on chromosome nine under wild type, STAG1-RNAi, and STAG2-RNAi conditions ( Figure 6B) was simulated as a polymer of length 2000, each monomer representing 1 kb of chromatin, for a total of 850000 times steps. A total of 480 replicated simulations were performed and aggregated into a single contact map. All simulations contained an average of three cohesin complexes actively extruding chromatin to form loops. The probability for one end of the extrusion complex to halt at a particular locus was derived from CTCF ChIP-seq data in HeLa cells (ENCODE phase 2, Broad Institute, file ENCFF000BAN), normalized to a probability between 0 and 1. Each halted end of an extrusion complex also had a 0.05% probability to continue sliding at each time step (halting lifetime of~2000 time steps). In addition, extrusion complexes have had predefined average lifetime on chromatin, and changes in STAG1 and STAG2 levels were simulated by modulating the extrusion lifetime. Wild type extrusion complexes had an average lifetime of 5000 timesteps (0.02% chance of dissociating at each time step) while STAG1-RNAi had 5-fold higher lifetimes and STAG2-RNAi had 2-fold lower lifetimes.
Simplified simulations of short, medium, or long extrusion lifetimes ( Figure 6A) were performed similarly. Chromatin was represented as polymers of length 1000 containing an average of three extrusion complexes. The two outermost CTCF binding sites had halting probabilities of 0.9 while the inner four binding sites had halting probabilities of 0.4. Halted extrusion ends had a 0.01% probability to continue sliding. Short, medium, and long extrusion lifetimes were modeled as complexes with lifetimes of 312.5, 625, and 10000-time steps respectively.

ChIP-seq peak calling and calculation of peak overlaps
Peaks were called by the MACS algorithm version 1.4.2 , using a P-value threshold of 1e-10 and by using sample and input read files. We identified sites of overlapping peaks between different conditions as well as between SMC3 and CTCF peaks using the MULTOVL software (Aszó di, 2012). We applied an inclusive type of overlap display ('union'), in which coordinates of overlapping peaks are merged into one common genomic site.

Hi-C library preparation
We generated a total of 14 in situ Hi-C libraries from our RNAi experiments (Supplementary file 1). Libraries generated by using MboI enzyme were done as described in Rao et al. (2014) without modification. In brief, the in situ Hi-C protocol involves crosslinking cells with formaldehyde, permeabilizing nuclei with detergent, digesting DNA overnight using a 4-cutter restriction enzyme, filling in 5'-overhangs while incorporating a biotinylated nucleotide, ligating newly blunted ends together, shearing DNA, capturing biotinylated ligation junctions with streptavidin beads, and analysing the resulting fragments with paired-end sequencing. All the libraries generated with the 6 bp cutter Hin-dIII were performed as in Wutz et al. (2017).

Hi-C data processing
All Hi-C libraries were sequenced with 150 bp paired-end reads and the resulting data was processed using Juicer (Durand et al., 2016b;Rao et al., 2014). These data were aligned against the hg19 reference genome. All contact matrices used for analysis were Knight-Ruiz or Vanilla-Coverage normalized with Juicer.
Loops were annotated in our RNAi experiments using hiccups (Durand et al., 2016a;Rao et al., 2014). Default parameters as described in Durand et al. (2016a); Rao et al. (2014) were used to call loops at 5 kb and 10 kb resolutions and merged as described in Rao et al. (2014). Domains were annotated in our RNAi experiments using Arrowhead (Durand et al., 2016a;Rao et al., 2014). Domains were called at 5 kb and 10 kb resolutions using default parameters and merged.